Comparative study of methods for human performance prediction using electro-encephalographic data

Abstract

The authors present here a comparative study of methods tackling the problem of predicting a person's quick or late response in an oddball experiment using their EEG data. The methods studied come from the related area of human performance monitoring (HPM) and rely on the use of kernel principal component analysis (KPCA), linear principal component analysis (LPCA), or time features combined with a support vector machine (SVM) or a Gaussian classifier. The results show the consistent superiority of the kernel PCA features, whereas SVM is marginally better than the Gaussian classifier. The classification rates produced with this combination of type of feature and classifier are moderate but they are significantly better than random for all subjects. This is important because it indicates that prediction of a person's performance using their EEG data is up to a certain extent feasible. This is a strong indication that early event related potential (ERP) components are related to brain's discrimination processes and are correlated with the reaction time in an oddball experiment.